Professor Kim Jeongho's lab envisions a super agent directing 100 specialized agents
Automating iterative design and verification... expected to shorten development time

"In the future, the competitiveness of engineers and developers will be determined not merely by their ability to use AI, but by their ability to design a range of specialized AI agents themselves and make them work together seamlessly."
KAIST Professor Kim Jeongho

KAIST Professor Kim Jeongho

View original image

A new approach to semiconductor development has been proposed in which AI agents handle everything from high-bandwidth memory (HBM) design and simulation to performance optimization. Instead of engineers operating design tools directly, AI repeatedly analyzes and refines designs to find the optimal solution. As semiconductor design grows more complex, this approach is drawing attention as a new way to reduce development time and costs.


On November 11, Professor Kim Jeongho of KAIST's School of Electrical Engineering presented the development direction for "Tera AI Agents," a system designed to automate semiconductor design, in a keynote address at the online "TERA AI Agent Day." The plan is to have a super agent direct 100 specialized agents.


The idea is to use AI once again in the research and development of HBM, which has supported advances in AI. The goal is to accelerate development and improve design quality by referencing specialized knowledge through retrieval-augmented generation (RAG) and using past design experience and work records.


Professor Kim's lab, Tera Lab, is developing "Alsipiti," an AI agent that analyzes and optimizes signal, power, and thermal integrity (SI, PI, and TI). The signal analysis agent examines signal distortion during data transmission and channel characteristics, while the power analysis agent optimizes the power distribution network (PDN) and capacitor placement. The thermal analysis agent assesses heat generation and cooling structures. If a design fails to meet the required standards, AI revises it and repeats the simulation. Engineers make the final decisions on design approval and mass production.


Experimental results demonstrating the potential of design automation were also presented. The research team compared two approaches to implementing a genetic algorithm for power distribution network optimization: one developed entirely by a person and the other developed in collaboration with ChatGPT. The two approaches delivered nearly identical optimization performance. This means development can be faster while achieving similar performance.



Professor Kim emphasized, "Semiconductor design and research and development are highly complex engineering fields in which diverse specialized areas, including circuits, packaging, and signal, power, and thermal integrity, are interconnected. The trend toward having AI agents share and collaborate on these specialized tasks will bring fundamental changes to semiconductor research and development."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing